2021.Q2: Time Reversal and the Maximum of Brownian Motion

2021 Probability Prelim Exam (PDF)

Introduction

This problem tests several core ideas about Brownian motion that often appear implicitly in prelim problems:

  • Understanding time-reversal of Brownian motion and recognizing when it produces another Brownian motion in distribution.
  • Working comfortably with Gaussian processes, especially using mean and covariance to identify distributions.
  • Connecting pathwise identities (almost sure statements) with distributional identities, especially involving extrema.
  • Correctly invoking the reflection principle and understanding what it actually gives.

Main tools likely needed:

  • Covariance structure of Brownian motion.
  • Characterization of Gaussian processes by mean and covariance.
  • Continuity of Brownian sample paths.
  • Reflection principle for Brownian motion.

Pattern recognition cue:
If you see expressions like $B_{T-t} - B_T$, think “time reversal + centering,” and if you see maxima like $B_t^*$, expect the reflection principle to appear.


Problem Statement (verbatim)

Problem 2.
In this problem ${B_t, t \ge 0}$ represents a Standard Brownian Motion (SBM) defined on its canonical space
$(\Omega = C[0, \infty), {\mathcal{F}t}{t \ge 0}, \mathcal{F}, P)$.

Let
\(B_t^* = \max_{0 \le s \le t} \{B_s\}, \quad t \ge 0\) (“trailing maximum” of SBM). Observe that with this notation the reflection principle states: \(B_t^* = |B_t| \quad \text{in distribution, for } t \ge 0.\) Let $T > 0$ be fixed.

a. Let $Y_t = B_{T-t} - B_T,\; 0 \le t \le T.$
Calculate $\operatorname{Cov}(Y_t, Y_s),\; 0 \le t, s \le T.$

b. Prove that ${Y_t, 0 \le t \le T} = {B_t, 0 \le t \le T}$ in distribution.

c.
(i) Prove that $Y_T^* = B_T^* - B_T,$ a.s.
(ii) Prove that $|B_T| = B_T^* - B_T$ in distribution.


Solutions

Part (a) Covariance computation

Claim.
For $0 \le s,t \le T$, \(\operatorname{Cov}(Y_t, Y_s) = \min(s,t).\)

Proof.
By definition, \(Y_t = B_{T-t} - B_T, \qquad Y_s = B_{T-s} - B_T.\) Thus \(\operatorname{Cov}(Y_t,Y_s) = \operatorname{Cov}(B_{T-t}-B_T,\; B_{T-s}-B_T).\) Expanding, $$ \begin{aligned} \operatorname{Cov}(Y_t,Y_s) &= \operatorname{Cov}(B_{T-t},B_{T-s})

  • \operatorname{Cov}(B_{T-t},B_T)
    &\quad - \operatorname{Cov}(B_{T-s},B_T)
  • \operatorname{Cov}(B_T,B_T). \end{aligned} \(Using $\operatorname{Cov}(B_a,B_b)=a\wedge b$,\) \begin{aligned} \operatorname{Cov}(Y_t,Y_s) &= (T-t)\wedge(T-s) - (T-t)\wedge T - (T-s)\wedge T + T. \end{aligned} \(If $s \le t$, this simplifies to\) (T-t) - (T-t) - (T-s) + T = s. \(Similarly, if $t \le s$, the result is $t$. Hence\) \operatorname{Cov}(Y_t,Y_s) = \min(s,t). $$

Conclusion.
The covariance structure of $(Y_t)$ matches that of standard Brownian motion.

Key Takeaways.

  • Brownian covariance: $\operatorname{Cov}(B_s,B_t)=s\wedge t$.
  • Always expand covariance linearly before simplifying.
  • Time-reversed increments often collapse cleanly after cancellation.

Part (b) Equality in distribution of processes

Claim.
\(\{Y_t, 0 \le t \le T\} \stackrel{d}{=} \{B_t, 0 \le t \le T\}.\)

Proof.
First, \(E[Y_t] = E[B_{T-t} - B_T] = 0 - 0 = 0 = E[B_t].\) From part (a), \(\operatorname{Cov}(Y_s,Y_t) = s \wedge t = \operatorname{Cov}(B_s,B_t).\) The process $(Y_t)$ is Gaussian since it is a linear transformation of the Gaussian vector $(B_{T-t}, B_T)$. A centered Gaussian process is completely determined by its covariance function. Therefore, $(Y_t)$ and $(B_t)$ have the same finite-dimensional distributions, hence the same distribution as processes on $[0,T]$.

Conclusion.
$(Y_t)$ is a standard Brownian motion in distribution.

Key Takeaways.

  • Linear transformations of Gaussian vectors remain Gaussian.
  • For Gaussian processes, mean + covariance completely determine the law.
  • “Equality in distribution of processes” means equality of all finite-dimensional distributions.

Part (c)(i) Pathwise identity for the maximum

Claim.
\(Y_T^* = B_T^* - B_T \quad \text{a.s.}\)

Proof.
By definition, \(Y_T^* = \max_{0 \le s \le T} Y_s = \max_{0 \le s \le T} (B_{T-s} - B_T).\) Since $B_T$ is constant in $s$, \(Y_T^* = \left(\max_{0 \le s \le T} B_{T-s}\right) - B_T.\) Let $u = T-s$. As $s$ runs from $0$ to $T$, $u$ runs from $T$ down to $0$, so \(\max_{0 \le s \le T} B_{T-s} = \max_{0 \le u \le T} B_u = B_T^*.\) Therefore, \(Y_T^* = B_T^* - B_T \quad \text{a.s.}\)

Conclusion.
The maximum of the time-reversed process has a simple pathwise relationship to the original maximum.

Key Takeaways.

  • Maxima commute with time-reversal when paths are continuous.
  • Be careful with change-of-variables in maxima: track bounds explicitly.
  • This part is almost entirely deterministic once paths are fixed.

Part (c)(ii) Distributional identity via reflection

Claim.
\(|B_T| \stackrel{d}{=} B_T^* - B_T.\)

Proof.
From part (b), $(Y_t){0 \le t \le T} \stackrel{d}{=} (B_t){0 \le t \le T}$. Since Brownian motion has continuous sample paths, the maximum over $[0,T]$ is determined by the path, and hence \(Y_T^* \stackrel{d}{=} B_T^*.\) From part (c)(i), we have the almost sure identity \(Y_T^* = B_T^* - B_T,\) so \(B_T^* - B_T \stackrel{d}{=} B_T^*.\) By the reflection principle, \(B_T^* \stackrel{d}{=} |B_T|.\) Combining these gives \(B_T^* - B_T \stackrel{d}{=} |B_T|.\)

Conclusion.
The reflected maximum minus the terminal value has the same law as the absolute value of Brownian motion.

Key Takeaways.

  • Reflection principle: $B_T^* \stackrel{d}{=} B_T $.
  • Continuity of paths allows extrema to inherit equality in distribution.
  • Distinguish carefully between almost sure identities and distributional equalities.

Master Key Takeaways

  • Time-reversed Brownian motion (properly centered) is Brownian motion in distribution.
  • For Gaussian processes, matching covariance and mean is enough.
  • Almost sure identities can be combined with distributional identities to produce new laws.
  • Continuity is the silent hypothesis that justifies passing from finite-dimensional distributions to maxima.

Cheat Sheet Entries to Extract

  • Time-reversal trick: $B_{T-t}-B_T$ has the same law as $B_t$ on $[0,T]$.
  • Gaussian process ID: Mean $=0$ + covariance $= s\wedge t$ ⇒ standard Brownian motion.
  • Reflection principle: $B_t^* \stackrel{d}{=} B_t $.
  • Maxima + continuity: For continuous processes, equality in distribution passes to suprema.

Notes on My Original Work

  • Strong: Correct covariance computation and correct high-level strategy throughout.
  • Needed patching: Clarifying the Gaussian-process argument in part (b).
  • Main fix: Explicitly invoking continuity (rather than abstract “functionals”) when passing to maxima.
  • Symbol errors in the handwritten version (e.g., $ B_T^* $) were corrected for clarity.

Comments